waste-predictor 4.0.0

Production-grade machine learning system for industrial waste prediction

Waste Prediction Module V4

A production-grade machine learning system for predicting industrial waste compositions based on production volume and environmental parameters.

📊 Model Performance

Metric	Score
R² Score	0.98
MAE	1,607
RMSE	3,092
CV R² (5-Fold)	0.974 ± 0.005

Per-Target R² Scores

Waste Type	R²
Total_Waste_kg	0.96
Solid_Waste_Limestone_kg	0.98
Solid_Waste_Gypsum_kg	0.99
Solid_Waste_Industrial_Salt_kg	0.98
Liquid_Waste_Bittern_Liters	0.99
Potential_Epsom_Salt_kg	0.98
Potential_Potash_kg	0.99
Potential_Magnesium_Oil_Liters	0.98

🚀 Quick Start

Installation

pip install -r requirements.txt

Training

python train_v4.py

Prediction

from predict_v4 import WastePredictorV4

# Load model
predictor = WastePredictorV4.load('waste_predictor_v4.pkl')

# Make prediction
result = predictor.predict(
    production_volume=50000,
    rain_sum=200,
    temperature_mean=28,
    humidity_mean=85,
    wind_speed_mean=15,
    month=6
)

print(result)
# {'Total_Waste_kg': 125000.5, 'Solid_Waste_Limestone_kg': 12500.0, ...}

Quick Prediction Function

from predict_v4 import quick_predict

result = quick_predict(
    production_volume=50000,
    rain_sum=200,
    temperature_mean=28,
    humidity_mean=85,
    wind_speed_mean=15,
    month=6
)

Batch Prediction

import pandas as pd
from predict_v4 import WastePredictorV4

predictor = WastePredictorV4.load('waste_predictor_v4.pkl')

# Create input DataFrame
data = pd.DataFrame({
    'Month': [1, 6, 12],
    'production_volume': [30000, 70000, 50000],
    'rain_sum': [300, 0, 250],
    'temperature_mean': [26, 28, 26],
    'humidity_mean': [100, 80, 98],
    'wind_speed_mean': [20, 12, 18]
})

# Get predictions
results = predictor.predict_batch(data)
print(results)

📁 Project Structure

local-module/
├── data/
│   └── training/
│       └── training.csv          # Training dataset (312 samples)
├── train_v4.py                   # V4 training (PRODUCTION - R²=0.98)
├── predict_v4.py                 # V4 inference module
├── waste_predictor_v4.pkl        # Trained model checkpoint
├── waste_predictor_v4_metadata.json
├── train_v3.py                   # V3 training (Neural network)
├── model_v2.py                   # Enhanced neural network model
├── feature_engineering.py        # Feature engineering pipeline
├── requirements.txt              # Dependencies
└── README.md

🔧 Input Features

Feature	Description	Range
production_volume	Production volume	0 - 200,000
rain_sum	Total rainfall (mm)	0 - 1,000
temperature_mean	Average temperature (°C)	0 - 50
humidity_mean	Average humidity (%)	0 - 100
wind_speed_mean	Average wind speed (km/h)	0 - 50
month	Month of year	1 - 12

📤 Output Predictions

Output	Description
Total_Waste_kg	Total waste produced (kg)
Solid_Waste_Limestone_kg	Limestone solid waste (kg)
Solid_Waste_Gypsum_kg	Gypsum solid waste (kg)
Solid_Waste_Industrial_Salt_kg	Industrial salt waste (kg)
Liquid_Waste_Bittern_Liters	Bittern liquid waste (L)
Potential_Epsom_Salt_kg	Potential Epsom salt byproduct (kg)
Potential_Potash_kg	Potential Potash byproduct (kg)
Potential_Magnesium_Oil_Liters	Potential Magnesium oil (L)

🧠 Model Architecture (V4)

Weighted Ensemble of 3 Model Types:

1. XGBoost Gradient Boosting (weight ~33%)

   • 500 estimators, max_depth=6
   • Per-target models with log-transformed outputs

2. Stacked Ensemble (weight ~34%)

   • Level 0: XGBoost + LightGBM + Random Forest + GBR
   • Level 1: Ridge regression meta-learner
   • 5-fold stacking with passthrough

3. Deep Neural Network (weight ~33%)

   • Architecture: 256 → 512 → 256 → 128 with skip connections
   • GELU activation, BatchNorm, Dropout
   • Cosine annealing LR schedule

Feature Engineering (30+ features):

• Log/sqrt/squared production transforms
• Cyclical month encoding (sin/cos)
• Weather condition indices (wet, dry, evaporation)
• Production × weather interactions
• Domain-driven ratio features

📈 Performance Comparison

Version	R² Score	MAE	Key Technique
V1 (Original)	0.47	7,873	Simple MLP
V3	0.77	5,575	Log transform + Feature eng
V4 (Production)	0.98	1,607	XGBoost + Stacked + DNN Ensemble

🔄 Retraining

To retrain the model with new data:

1. Add new data to data/training/training.csv
2. Run training:

   python train_v4.py
   

3. Model will be saved to waste_predictor_v4.pkl

📝 License

MIT License